Environment:Protectai Llm guard Python Runtime Dependencies

Knowledge Sources	LLM Guard pyproject.toml
Domains	NLP, Security, Infrastructure
Last Updated	2026-02-14 12:00 GMT

Overview

Python 3.10+ environment with PyTorch, HuggingFace Transformers, Presidio NLP, and tiktoken for running LLM Guard scanner pipelines.

Description

This environment provides the core runtime for LLM Guard, a security framework for Large Language Models. It requires Python 3.10-3.12 with PyTorch for deep learning inference, HuggingFace Transformers for model loading and text classification pipelines, Presidio for PII entity recognition and anonymization, tiktoken for OpenAI-compatible token counting, and several supporting libraries for text processing. The framework auto-detects CUDA, MPS, or CPU devices at startup via torch.cuda.is_available() and torch.backends.mps.is_available().

Usage

Use this environment for any LLM Guard scanner operation, including the core library (llm_guard package), benchmarking, and integration examples. This is the mandatory prerequisite for all scanner implementations including Anonymize, PromptInjection, Toxicity, TokenLimit, Relevance, and others.

System Requirements

Category	Requirement	Notes
OS	Linux, macOS, Windows (WSL recommended)	Tested on Ubuntu in CI
Python	3.10, 3.11, or 3.12	`requires-python = ">=3.10,<3.13"`
Hardware	CPU (minimum), NVIDIA GPU (recommended)	CUDA auto-detected; MPS supported on Apple Silicon
RAM	4GB minimum, 8GB+ recommended	Transformer models require significant memory

Dependencies

System Packages

Python 3.10+ runtime
git (for cloning and model downloads)

Python Packages

torch >= 2.4.0
transformers == 4.51.3
presidio-analyzer == 2.2.358
presidio-anonymizer == 2.2.358
tiktoken >= 0.9, < 1.0
nltk >= 3.9.1, < 4
faker >= 37, < 38
fuzzysearch >= 0.7, < 0.9
bc-detect-secrets == 1.5.43
json-repair == 0.44.1
regex == 2024.11.6
structlog >= 24

Credentials

No credentials are required for the core library. However, certain HuggingFace models (gated models like deberta-v3-small-prompt-injection-v2) require:

HF_TOKEN: HuggingFace API token with read access for gated model downloads.

Quick Install

# Install core package
pip install llm-guard

# If torch install fails, install separately first
pip install wheel
pip install torch>=2.4.0
pip install llm-guard --no-build-isolation

# For development
pip install "llm-guard[dev]"

Code Evidence

Device auto-detection from llm_guard/util.py:103-111:

@lru_cache(maxsize=None)  # Unbounded cache
def device():
    torch = cast("torch", lazy_load_dep("torch"))
    if torch.cuda.is_available():
        return torch.device("cuda:0")
    elif torch.backends.mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")

Lazy dependency loading with warning from llm_guard/util.py:114-131:

def lazy_load_dep(import_name: str, package_name: str | None = None):
    if package_name is None:
        package_name = import_name
    spec = importlib.util.find_spec(import_name)
    if spec is None:
        LOGGER.warning(
            f"Optional feature dependent on missing package: {import_name} was initialized.\n"
            f"Use `pip install {package_name}` to install the package if running locally."
        )
    return importlib.import_module(import_name)

Python version constraint from pyproject.toml:19:

requires-python = ">=3.10,<3.13"

Common Errors

Error Message	Cause	Solution
`ModuleNotFoundError: No module named 'torch'`	PyTorch not installed	`pip install torch>=2.4.0`
`LookupError: punkt_tab`	NLTK tokenizer data missing	Auto-downloaded on first use; ensure internet access
`Optional feature dependent on missing package: tiktoken`	tiktoken not installed	`pip install tiktoken`
`TOKENIZERS_PARALLELISM` warning	HuggingFace tokenizer threading conflict	Automatically set to `false` by Anonymize scanner

Compatibility Notes

CUDA GPUs: Auto-detected via torch.cuda.is_available(). Uses cuda:0 by default.
Apple Silicon (MPS): Supported via torch.backends.mps.is_available().
CPU-only: Full functionality available; ML-based scanners will run slower.
Python 3.9: Listed in some documentation but pyproject.toml requires >=3.10.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment